The Alignment Problem

How Can Machines Learn Human Values?

Brian Christian

• Publisher: Atlantic Books
• Year: 2021
• ISBN:
  9781786494320
• Topic: Informatique
• Subtopic: Sciences générales de l'informatique

'Vital reading. This is the book on artificial intelligence we need right now.' Mike Krieger, cofounder of Instagram Artificial intelligence is rapidly dominating every aspect of our modern lives influencing the news we consume, whether we get a mortgage, and even which friends wish us happy birthday. But as algorithms make ever more decisions on our behalf, how do we ensure they do what we want? And fairly?This conundrum - dubbed 'The Alignment Problem' by experts - is the subject of this timely and important book. From the AI program which cheats at computer games to the sexist algorithm behind Google Translate, bestselling author Brian Christian explains how, as AI develops, we rapidly approach a collision between artificial intelligence and ethics. If we stand by, we face a future with unregulated algorithms that propagate our biases - and worse - violate our most sacred values. Urgent and fascinating, this is an accessible primer to the most important issue facing AI researchers today.

INDEX

0–1 loss function, 354n47

3CosAdd algorithm, 316, 397n13

Abbeel, Pieter, 257, 258–59, 267–68, 297

Ackley, Dave, 171–72

actor-critic architecture, 138

actualism vs. possibilism, 234–40

effective altruism and, 237–38

imitation and, 235, 239–40, 379n71

Professor Procrastinate problem, 236–37, 379n61

actuarial models, 92–93

addiction, 135, 153, 205–08, 374n65

See also drug use

AdSense, 343n72

adversarial examples, 279–80, 387n8

affine transformations, 383n16

African Americans. See racial bias

Against Prediction (Harcourt), 78–79

age bias, 32, 396n7

AGI. See artificial general intelligence

Agüera y Arcas, Blaise, 247

AI Now Institute, 396n9

AI safety

artificial general intelligence delay risks and, 310

corrigibility, 295–302, 392–93n51

field growth, 12, 249–50, 263

gridworlds for, 292–93, 294, 295, 390n29

human-machine cooperation and, 268–69

irreversibility and, 291, 293

progress in, 313–14

reward optimization and, 368n56

uncertainty and, 291–92

See also value alignment

AIXI, 206–07, 263

Alciné, Jacky, 25, 29, 50

ALE. See Arcade Learning Environment

Alexander, Robert McNeill, 382n13

Alexnder VIII (Pope), 303

AlexNet

brittleness of, 279

complexity of, 100

development of, 21, 23–25

dropout and, 24, 285, 389n21

ensembles and, 285

feature visualization and, 108, 109–10

ground truth gaps and, 75

algorithms, ...